MasticatoryMuscleMusculomics

nnU-Net v2 model segmenting the four substructures of the lateral pterygoid muscle (right/left superior and inferior heads) on non-contrast head CT, packaged as a TotalSegmentator additional model. Companion code and documentation: the MasticatoryMuscleMusculomics GitHub repository.

Companion paper: Automated Masticatory Muscle Musculomics on Routine Head CT: Multicenter Reproducibility and Cerebrovascular Status Classification (under peer review; author and affiliation information anonymized for double-blind review).

Model details

• Task: 4-label segmentation of the lateral pterygoid substructures (1 right superior, 2 right inferior, 3 left superior, 4 left inferior)
• Input: non-contrast head CT (full field of view, no pre-cropping)
• Architecture: nnU-Net v2 ResidualEncoderUNet-L (nnUNetResEncUNetLPlans_torchres, 3d_fullres), custom trainer nnUNetTrainerLRSwapMirrorNoRotation300 (left/right label-aware mirroring, no rotation augmentation, 300 epochs; inference TTA excludes the left-right axis)
• Fold: fold_0; released weights are the benchmark checkpoint_best.pth, shipped as checkpoint_final.pth
• Training data: 264 non-contrast head CTs (train/val/test = 185/26/53, seed 42), multi-center cohort

Performance

Frozen internal test set, 53 cases: Dice 0.952 (95% CI 0.945–0.958), NSD@1mm 0.971, HD95 0.836 mm (case-level mean over the 4 labels, bootstrap 95% CI).

Usage

Download the weights zip and install it into the TotalSegmentator weights directory (the zip has the same structure as official TotalSegmentator weight packages; SHA256: 03f6d6dc796ea2a6d74af625ecde04ed5fbc4d3a5a00b3b3183b3305e784bdbc):

hf download frankzhang/MasticatoryMuscleMusculomics Dataset963_MasticatoryMuscleDetail.zip --local-dir .

# inside a clone of the companion GitHub repository
python install_model.py --zip Dataset963_MasticatoryMuscleDetail.zip

# then run inference
totalseg-masticatory -i ct.nii.gz -o out_dir --device gpu:0

The repository also contains musc_geometry/, the musculomics feature-extraction module (morphological and density/texture features) used in the companion study.

Citation

If you use this model, please cite TotalSegmentator and nnU-Net. The citation for the companion paper will be added upon publication.

License

MIT. The model is intended for research purposes only; it is not a medical device and has not been approved for clinical use.